Das Gupta Avishek, Sadek Zafar, Hossain Md Shakhawat, Toha Tarik Reza, Mondol Anupom, Habiba Sultana Umme, Mominul Alam Shaikh Md
Department of Textile Machinery Design and Maintenance, Bangladesh University of Textiles, Tejgaon, Dhaka-1208, Bangladesh.
Heliyon. 2024 Aug 23;10(17):e35931. doi: 10.1016/j.heliyon.2024.e35931. eCollection 2024 Sep 15.
Knit fabric is one of the dominating fabric types for wearable textile across the whole world and during the production of knit fabric, faults created by different reasons cause difficulties in the subsequent process. The existing fault detection process in the knitting industry is done manually by the human naked eye. To detect faults automatically, deep learning-based models are very efficient and can reduce the workload for fabric inspection. However, implementing a deep learning-based model for fabric fault detection needs a large number of images with different types of knit fabric faults from real-world scenarios to train and for better performance. Most of the existing fabric fault datasets are based on woven fabric, which cannot be used for training because the faults of woven fabric and knit fabrics are completely different. In this research, we propose a new dataset named ISL-Knit. It is a benchmark dataset for knit fabric faults collected from different knit dyeing industries in Bangladesh. Our dataset contains high-resolution images of grey fabric and dyed fabric annotated with 7 types of faults. To provide correct annotation, all images of faults are annotated and verified with the proper reference which can be a great tool for boosting up deep learning-based models for automatic knit fabric fault detection in the textile field. A comparison of the accuracy of the deep learning-based model by training with our proposed dataset and existing knit fabric fault dataset is also done. To develop an automatic fault detection system, we trained the YOLOv5 models with our dataset, considering different image sizes. For real-world implementation, the best models are selected considering mAP and inference time. YOLOv5 models are implemented in Raspberry Pi with a Raspberry Pi camera, a laptop with no dedicated GPU, and a laptop with a dedicated GPU with a web camera to observe the performance. Finally, an automated four-point inspection system is developed by calibrating the camera to generate the score representing the quality of fabrics. The implementation represents the feasibility of a deep learning-based model in knit fabric fault detection, considering the cost and accuracy.
针织面料是全球可穿戴纺织品中占主导地位的面料类型之一，在针织面料生产过程中，由不同原因产生的瑕疵会给后续加工带来困难。针织行业现有的瑕疵检测过程是由人工肉眼手动完成的。为了自动检测瑕疵，基于深度学习的模型非常高效，并且可以减少面料检测的工作量。然而，要实现一个基于深度学习的针织面料瑕疵检测模型，需要大量来自真实场景的、包含不同类型针织面料瑕疵的图像进行训练，以获得更好的性能。现有的大多数面料瑕疵数据集都是基于机织面料的，由于机织面料和针织面料的瑕疵完全不同，因此不能用于训练。在本研究中，我们提出了一个名为ISL-Knit的新数据集。它是一个从孟加拉国不同针织印染行业收集的针织面料瑕疵基准数据集。我们的数据集包含灰色面料和染色面料的高分辨率图像，并标注了7种瑕疵类型。为了提供正确的标注，所有瑕疵图像都经过了适当参考的标注和验证，这对于推动纺织领域基于深度学习的自动针织面料瑕疵检测模型来说是一个很好的工具。我们还对使用我们提出的数据集和现有针织面料瑕疵数据集进行训练的基于深度学习的模型的准确率进行了比较。为了开发一个自动瑕疵检测系统，我们使用我们的数据集对YOLOv5模型进行训练，考虑了不同的图像尺寸。为了在实际中应用，我们根据平均精度均值（mAP）和推理时间选择了最佳模型。YOLOv5模型在配备树莓派摄像头的树莓派、没有专用GPU的笔记本电脑以及配备网络摄像头的有专用GPU的笔记本电脑上实现，以观察其性能。最后，通过校准摄像头开发了一个自动四点检测系统，以生成代表面料质量的分数。考虑到成本和准确性，该实现展示了基于深度学习的模型在针织面料瑕疵检测中的可行性。
